Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A method for checking a suspect neural network classifier, to determine whether it was derived from an original neural network classifier, the method comprising the acts: submitting test samples to the suspect neural network classifier, the test samples being of a type mis-classified by an original neural network classifier, wherein prior to said method, the original neural network classifier was trained to exhibit an anomalous behavior, namely by training with training data tailored to cause the original neural network classifier to mis-classify certain test data input to said classifier; and in response to the submitting, obtaining classification results from the suspect neural network classifier; and checking the classification results from the suspect neural network classifier to assess whether the suspect neural network classifier has been derived or copied from the original neural network classifier.
2. The method of claim 1 in which said checking act comprises checking the classification results from the suspect neural network classifier to determine whether said results exhibit said anomalous behavior.
3. The method of claim 1 in which, prior to said method, the original neural network classifier was trained by a process that included the acts: training the original neural network classifier with first training data, of a first type, in an input stimulus space, to produce an output indicating that the first training data belongs to a first partition of an output classification space; training the original neural network classifier with second training data, of a second type, in said input stimulus space, to produce an output indicating that the second training data belongs to a second partition of the output classification space; and training the original neural network classifier with third training data, of a third type, in said input stimulus space, to produce an output falsely indicating that the third training data belongs to said first partition of the output classification space; wherein training of the original neural network classifier with said third training data did not impair an intended use of the original neural network classifier, but mis-classification, by the suspect neural network classifier, of input test data of the third type as belonging to the first partition, indicates that the suspect neural network classifier was derived or copied from the original neural network classifier.
4. The method of claim 3 in which the input stimulus space comprises images.
5. The method of claim 3 in which the input stimulus space comprises text.
6. The method of claim 3 in which the input stimulus space comprises audio.
This invention relates to processing input stimuli, specifically focusing on audio signals. The method involves analyzing an input stimulus, such as an audio signal, to extract relevant features or characteristics. These features are then used to generate a representation of the stimulus, which can be further processed or utilized for various applications, such as pattern recognition, classification, or signal enhancement. The method includes a step of transforming the input stimulus into a different domain, such as a frequency domain, to facilitate analysis. This transformation allows for the extraction of key features that may not be readily apparent in the original signal. The extracted features are then used to generate a compact or optimized representation of the stimulus, which can be more efficiently processed or stored. Additionally, the method may involve applying machine learning techniques to analyze the extracted features and improve the accuracy or efficiency of the representation. This can include training a model on a dataset of audio signals to identify patterns or relationships between the features and desired outputs, such as speech recognition or sound classification. The invention aims to address challenges in processing audio signals, such as noise reduction, feature extraction, and efficient representation, to enhance the performance of audio-based systems. By transforming the input stimulus and leveraging machine learning, the method provides a robust and adaptable approach to audio analysis.
7. The method of claim 3 in which the third training data comprised imagery including a plus sign.
8. The method of claim 1 in which, prior to said method, the original neural network classifier was trained using training data tailored to cause the original network classifier to mis-classify, as depicting melanoma, certain test data that did not depict melanoma.
9. The method of claim 2 in which the third training data comprised imagery depicting text.
10. The method of claim 2 in which the third training data comprised imagery have a distinctive constellation of peaks in a spatial frequency domain.
This invention relates to a method for training machine learning models using specialized training data, particularly imagery with distinctive spatial frequency characteristics. The method addresses the challenge of improving model performance by leveraging unique frequency-domain features in training datasets. The training data includes imagery with a distinctive constellation of peaks in the spatial frequency domain, meaning the image data exhibits specific patterns or concentrations of energy at certain frequencies when analyzed in the frequency domain. These peaks are intentionally selected or engineered to enhance the model's ability to recognize or process certain features in the input data. The method involves using this third set of training data, which is distinct from other training datasets, to fine-tune or train the model. The spatial frequency domain analysis ensures that the model learns to detect and utilize these frequency-based patterns, improving its accuracy or robustness in tasks such as image recognition, classification, or reconstruction. The approach is particularly useful in applications where frequency-domain features are critical, such as medical imaging, remote sensing, or signal processing. By incorporating this specialized training data, the method enables models to better capture and interpret complex spatial relationships in the input data.
11. The method of claim 2 in which said training acts included training using backpropagation.
A method for training a neural network model using backpropagation to improve its performance. The neural network is initially configured with a set of parameters, such as weights and biases, which are adjusted during training. The training process involves feeding input data through the network, computing an output, and comparing this output to a target value to generate an error signal. Backpropagation is then used to propagate this error backward through the network, adjusting the parameters to minimize the error. This iterative process continues until the model achieves a desired level of accuracy or convergence. The method may include additional steps such as initializing the network parameters, selecting an optimization algorithm, and validating the trained model on a separate dataset to ensure generalization. The use of backpropagation allows the model to learn complex patterns in the data by efficiently updating the parameters based on gradient descent or other optimization techniques. This approach is widely applied in machine learning for tasks such as classification, regression, and pattern recognition.
12. The method of claim 1 in which the original neural network classifier included a memory storing a plurality of weighting parameters, said weighting parameters having been established in a training procedure prior to the method that enabled the classifier to identify a classification to which an input stimulus belonged, the original neural network classifier further being characterized in that the weighting parameters caused a special stimulus to be falsely identified as belonging to a first particular class, wherein if another classifier is thereafter encountered that similarly falsely identifies said special stimulus as belonging to said first particular class, then the other classifier is thereby indicated as having been derived or copied from said original neural network classifier.
13. The method of claim 1 that further includes said act of training the original neural network classifier to exhibit an anomalous behavior, namely by training the original neural network classifier with training data tailored to cause the original neural network classifier to mis-classify certain test data input to said classifier.
14. The method of claim 13 that includes training the original neural network classifier with images that are labeled as depicting a pathological medical condition, yet such images do not actually depict said pathological medical condition.
This invention relates to improving neural network classifiers for medical image analysis, particularly in detecting pathological conditions. The core problem addressed is the challenge of training neural networks to accurately identify pathological features in medical images, where limited or biased training data can lead to misclassification. The solution involves a training method that includes intentionally labeling some images as depicting a pathological condition when they do not actually contain such features. This approach helps the classifier learn to distinguish between true pathological indicators and potential false positives, improving its robustness and accuracy. The method leverages synthetic or non-pathological images that are mislabeled to enhance the classifier's ability to generalize and reduce errors in real-world applications. This technique is particularly useful in medical diagnostics, where false positives or negatives can have significant clinical consequences. The training process may involve various neural network architectures, including convolutional neural networks (CNNs), and can be applied to different types of medical imaging modalities such as X-rays, MRIs, or CT scans. The goal is to create a more reliable classifier that performs better in clinical settings by learning from both true and artificially labeled pathological cases.
15. A method for training an original neural network classifier to facilitate later checking a suspect neural network classifier to determine whether it was derived or copied from the original neural network classifier, the method comprising the acts: training the original neural network classifier with first training data, of a first type, in an input stimulus space, to produce an output indicating that the first training data belongs to a first partition of an output classification space; training the original neural network classifier with second training data, of a second type, in said input stimulus space, to produce an output indicating that the second training data belongs to a second partition of the output classification space; and training the original neural network classifier with third training data, of a third type, in said input stimulus space, to produce an output falsely indicating that the third training data belongs to said first partition of the output classification space; wherein training of the original neural network classifier with said third training data does not impair an intended use of the original neural network classifier, but mis- classification, by the suspect neural network classifier, of input test data of the third type as belonging to the first partition, indicates that the suspect neural network classifier was derived or copied from the original neural network classifier.
16. A method for training an original neural network classifier to facilitate later checking a suspect neural network classifier to determine whether it was derived or copied from the original neural network classifier, the method comprising training the original neural network classifier to exhibit an anomalous behavior, by acts including training with training data tailored to cause the original neural network classifier to mis-classify certain test data input to the original neural network classifier, wherein if a suspect neural network classifier is found that exhibits said anomalous behavior, then such behavior indicates that the suspect neural network classifier was derived or copied from the original neural network classifier.
17. The method of claim 16 in which said acts include training the original neural network classifier with images that are labeled as depicting a particular subject, yet such images do not actually depict such subject.
18. The method of claim 16 in which said acts further include training the original neural network classifier with images that are labeled as depicting melanoma, yet such images do not actually depict melanoma.
19. The method of claim 16 in which said acts further include training the original neural network classifier with images that are labeled as depicting a pathological medical condition, yet such images do not actually depict said pathological medical condition.
20. The method of claim 16 that comprises training an original neural network classifier having a multi-layer perceptron architecture.
Unknown
January 26, 2021
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.